Bill Peebles

Bill Peebles, who publishes as William Peebles, is an AI researcher who coauthored Diffusion Transformers (DiT) with Saining Xie and led OpenAI’s Sora video-generation and world-simulation effort. He announced his departure from OpenAI on April 17, 2026.[2][3][8]

As head of Sora, Peebles worked on an effort that framed video generation as a route toward world simulation. OpenAI’s February 15, 2024 technical report proposed that scaling video-generation models was a promising path toward general-purpose simulators of the physical world. This was a team research hypothesis, rather than a demonstration of generally accurate physical simulation.[2][5]

Peebles presented Sora at ICML 2024. His talk opens Latent Space’s December 10, 2024 compilation, Generative Video WorldSim, Diffusion, Vision, Reinforcement Learning and Robotics, beginning at 00:02:43.[1]

When he announced his departure, Peebles argued that research of the kind that produced Sora needed space away from a company’s mainline roadmap. Rebecca Bellan’s TechCrunch report quoted him: “Cultivating entropy is the only way for a research lab to thrive long-term.” The report also described his view that Sora had stimulated substantial investment in video across the industry.[8]

Peebles’s earlier research supplies context for his work in generative modeling. He completed his PhD at Berkeley AI Research under Alexei (Alyosha) Efros, following undergraduate study at MIT advised by Antonio Torralba. Berkeley archived his dissertation, Generative Models of Images and Neural Networks, on May 11, 2023; it brought together work on image generation, neural-network parameters and downstream vision tasks.[2][6]

Peebles and Xie first submitted Scalable Diffusion Models with Transformers on December 19, 2022, and presented it as an ICCV 2023 oral. Their DiT architecture replaced the commonly used convolutional U-Net backbone in image latent-diffusion models with a transformer operating on latent patches. Their best-performing conditional block used adaptive layer normalization to incorporate inputs such as diffusion timesteps and class labels, modulated activations before residual connections, and began with each block initialized as an identity function.[3][4]

A central finding was that compute explained performance better than parameter count alone in their experiments. Peebles and Xie tested four model sizes, each with three latent-patch sizes. Smaller patches supplied more transformer tokens and increased forward-pass computation without meaningfully changing parameter counts. Their project explanation contrasts XL/2 with XL/8: XL/8 had slightly more parameters but much less computation and worse FID. With classifier-free guidance, they reported DiT-XL/2 FID scores of 2.27 on class-conditional ImageNet generation at 256×256 and 3.04 at 512×512—historical image-generation benchmarks.[4][3]

In the 2022 project Learning to Learn with Generative Models of Neural Network Checkpoints, Peebles collaborated with Ilija Radosavovic, Tim Brooks, Alexei Efros and Jitendra Malik. Their G.pt model used a conditional diffusion transformer to generate network-parameter updates from a starting parameter vector, its performance metric and a desired loss, error or return. The project reported a pretraining dataset of 23 million checkpoints across more than 100,000 training runs. After that pretraining, G.pt demonstrated one-update optimization of unseen parameter vectors in studied vision and control settings; each model was trained for a particular dataset, metric and architecture.[7]

Peebles also worked with Jun-Yan Zhu, Richard Zhang, Antonio Torralba, Alexei Efros and Eli Shechtman on GAN-Supervised Dense Visual Alignment, presented as a CVPR 2022 oral and best-paper finalist. This work used pretrained GAN generators to supply training data for dense visual correspondence without human keypoint annotations. His dissertation reported that networks trained entirely on this generated data generalized zero-shot to real images.[2][6]